Researches On Video Anomaly Detection Model Based On Visual Attention

In recent years,with the occurrence of various public security,disaster detection,social anti-terrorism and so on,intelligent video anomaly detection technology plays an increasingly important role in public places such as squares,railway stations,banks and airports.Traditional video anomaly detection methods rely on manual analysis and video processing,but massive video data greatly increases the workload.They are not only inefficient but also easy to miss suspicious targets and abnormal events.The research of surveillance video anomaly detection can meet the requirements of automatic real-time alarm in video data analysis applications.For video anomaly detection applications,they combine technologies from multiple domains,such as image processing,pattern recognition and artificial intelligence,to understand the visual scenes by analyzing and processing video sequences.By using visual saliency detection technology,the computer can pay more attention to monitoring important visual regions in the video sequences while filtering out relatively unimportant visual regions,thereby allocating more computing resources to important regions,reducing computational cost and speeding up processing.This paper deeply studies the video saliency detection and video anomaly detection,and proposes new computational models for video saliency detection and anomaly detection.The main innovations of this paper are as follows:(1)Based on the powerful feature representation ability of deep convolutional networks,this paper proposes a new video saliency detection model by deep semantics and spatiotemporal cues,considering the intrinsic relationship between top-down and bottom-up saliency features.The model first extracts the top-down semantic features,then combines them with the bottom-up spatiotemporal salient features for the final saliency prediction.In order to effectively combine these features,we input them into the deconvolutional network model for feature sharing learning between semantic features and spatiotemporal cues to calculate the final saliency map.In addition,we propose a new loss function based on the Gaussian-like function and the L2-norm regularization term to optimize the proposed model.The experimental results show that,compared with existing video saliency detection methods,the prediction performance of the saliency detection model proposed in this paper is better.(2)In this work,we study the anomaly detection modeling problem in video sequences from the following two aspects.First,we constructed a new large-scale anomaly detection dataset as a benchmark dataset for video anomaly detection and classification.The data set includes 1200 normal and anomalous video sequences,including 12 anomaly types,including crash,fire,violence,et al.We also labeled video-level data(abnormal/normal video anomaly types)and frame-level data(abnormal/normal video frames).Second,we propose an effective anomaly detection calculation model by learning the local and global spatiotemporal context features.Experimental results show that this method is superior to existing anomaly detection methods.